Latest / The Bleeding Edge Podcast / Your AI Models Should Be Talking to Each Other
Transcript
- Ralph: ever with me in the studio today. Emil Ozier. Emil, welcome to the edge. Em: Great to be back, Rolf. Ralph: So it's been a staggering week. I think we've had the most new models dropped in a week in in the last 12 months. So a bit of a a record. We always talk about how things are speeding up, and this week they really sped up. So let's jump into the news in 60 seconds, and I think I've actually got this down to a fine art now. So Let's go. Headlines of the week. Four frontier models shipped in seven days. And we'll talk about each of those in turn. The big news is the price of a token fell by half. And the people who have been building the LLMs have pitched up at the UN Security Council to explain themselves. By all accounts, it's the first time that they've gone in and briefed the council directly. So talking about those releases, OpenAI released GPT 6 Seoul. Sol is their big flagship. it's the equivalent of Fable 5. And they also launched Luna, which is their budget pricing. and obviously at the same time, and this is really great news, they cut API API pricing by roughly 50%. And so if you're somebody that accesses your LLMs. via an API, then it's just got fifty percent cheaper, which for most developers is a reason to go and buy champagne. I certainly Em: Yeah. Ralph: am. And so Anthropic at the same time released Claude Opus five point five which I know Emil both of us have been very happy with because the ultra verbose Opus five was starting to piss me off. I was having to do so much reading And thankfully, 5.5 has cut that down. It is a lot more to the point and lightning fast. So I'm really enjoying that. Google pushed Gemini 3.8 live into voice. Alibaba shipped Quinn 3.8 Omni. And obviously in the same week, we've seen Sam and Dario and Clement Delange. Clement is the guy from Hugging Face. Or one of the guys from Hugging Face, Nvidia's just bought them as we kind of briefed everybody in in a couple of episodes ago. but anyway, all three of them went to the UN Security Council. and of course, Dario has been calling for a slowdown in AI, and everybody else has said no, including Trump, who said he wants the US policy left exactly where it is. and I think Xi Jinping is actually going to be at the White House in the next week. and no doubt they'll be talking about how they actually collaborate on the progress of of AI. Hopefully I'll get to a good deal. but I guess that is probably just more hopeful than reality. So the other big news is that Nvidia has actually launched Soul Pie. Which cuts coding agent traffic by up to 49%. So imagine not only are is your APIs, your API traffic with OpenAI costing you 50% less, but if you actually implemented Sol Pi, you could cut that by another 49%. And so all of this means that suddenly really capable LLMs have gotten very, very accessible. And the ROI on so many use cases has completely flipped. And to just illustrate this point, we've actually got a little model here. So basically a compaction plugin able to squeeze one million tokens to 86k in a second. So that's a a pretty big compaction and what it means is that you've now got an example of a browser agent booking a flight forget this it did it in seven seconds by the way And it cost 0.0039 cents. So that's dollar cents, which is a lot less than it would have cost your PA to go away and book your flight, and a lot faster too. And so a really big deal here just in terms of capability. that was probably a bit longer than 60 seconds, but a jam-packed week nonetheless, with four big models dropping, really exciting development from Nvidia. And everything just getting cheaper and the model economics suddenly starting to make sense in a very different way. What do you make of all of that, Emil? Em: Totally agree. we we share the same opinion on Opus. We're just both loving it. and above the fact that it's less verbose, I was just finding like the same frustrations as you. The language not only was it verbose, it was difficult to read, it didn't make sense, it was very technical, and you do that all day long, it's very disengaging. It it it reduces your your willingness to to wanna be engaged and fully aware on all those turns and I'm finding, you know, a new expressed interest when when I'm developing, I'm enjoying the output. It makes sense. It's a value add. It's helping me steer further direction. So it's not just efficient, it's helpful and it's keeping me interested. So fully welcomed. And on on the frontier aspect, couldn't agree more, Rolf. Just everywhere we look there's efficiencies. from everything from compaction, which you've mentioned, there's massive cash savings. all the big frontier lads are getting smarter about how they do cash reads. And that's really significant, especially with developments. we've got a lot of clever routing logic coming through. we've got the likes of things called mono threads, which is also bringing a lot of efficiencies. And what I find really interesting about this is when we initially started with AI using the chatbots, we got very familiar with the user interface on chatting interns. And I thought I thought that was going to develop into something else. It was it was a temporary solution as as we got more mature. But what's I'm finding quite interesting is it feels like the labs have decided to keep that interface and instead of changing it, they they're bringing the intelligence into it in the back end. And by that what I mean is now the likes of like in Claude, if you open up a conversation chat and you need to do something more complex, what it actually does underneath the bonnets is it spins up additional agents who create threads on conversations that connect to your main thread. So you have to be less concerned with saving your memory and worrying about when your context window is getting too large and how how to handle that handover. That's all happening in the background now and w with all that compaction and efficiency. And it's really fun to use. and in addition to that, Claude have also introduced new projects which which support everything that I'm speaking about. And we haven't even covered Jev yet which is a article all in itself we'll we'll leave for a another podcast. But yep, in total agreement. Excited. Ralph: Yeah, I think the whole thing has suddenly started to mature with kind of, you know, multiple developments coming together. And I just want to touch on a couple of things that you mentioned there because probably one of the biggest transformations that we are really kind of experiencing at the moment is that if you are a developer, you have probably stopped coding. you're maybe kind of coding as you know a kind of pastime like you would do Sudoku on a on a weekend. and this is kind of coming from everywhere. So I was listening to a wonderful interview that Corey Knowles from the Neuron great newsletter on AI did with Chen Goldberg, who's the executive vice president of product and engineering at Coreweave. And Core Weave have just picked up a whole bunch of Vera Rubin GPUs from NVIDIA. So they've basically built, you know, what she calls a supercomputer, the woman's smile on her face, the level of excitement that she had at having this kind of power in her hands. And they're building kind of cloud compute AI type solutions for for private clients. but what she was saying was, you know, actually the the big transformation is that you're not coding anymore. You are spending your time reviewing. So we are all now the the human gatekeepers, the true bottleneck is in the review process. And and that's what we've become. And if you're a developer, and I've actually read a lot of blog posts from developers that have said that's it, I'm done. You know, this isn't what I signed up for. But of course, for the rest of us that weren't developers to start with, I don't want to code, I want to build solutions, I want to solve big problems. I want to do it fast. And so this age of AI has really unlocked that. And I think, you know, maybe it's easier if you were a developer and you loved coding and you now can't spend your time doing that. Maybe it's bad. But for the rest of us, it's awesome. And the only thing that got in the way was you know, this kind of verbose model that we've just recently seen with Opus V and I I c you know, exactly what you said, it became a labor. I didn't want to do it anymore. I had too much to read. And so with 5.5 kind of making it fun again, you know, it got really interesting. And I think this is the trick. This is where the transformation is happening. there's so much changing, but the way we ha are spending our time, we've been literally glued. To various chat outputs, having to read absolutely everything to understand what's going on, to make the right decisions. And I found myself with 5.5 losing the will to live. well, with five. And now with 5.5, you know, I feel definitely reborn and the fun has returned. And so hopefully we kind of keep that going. but I just thought that was just such an interesting shift that we've seen, you know. I don't think, I don't know. Do you know any teams that are still Hand crafting, hand cutting code. Em: None. And I just want to also add to that, it's not just the efficiency, Rolf. being giving it quite a bit of thought. So when you have those big verbose responses, because you're exhausted and is pulling the will from you, you you don't have as a keen eye on everything. And that can lead to mistakes too. And also just Claude in itself, when it marks its own work and it uses that reasoning logic on each of the turns. It's is reducing the chances of hallucinations, which invariably creates better safety all around for the product, but you working on yourself engaged on it as well as the model using its intelligence against that context. And the other aspect is we're hearing a lot in the news, we're hearing a lot about like the the big slowdown on AI, but what I'm finding really refreshing is they are slowing down on Frontier on the cutting edge stuff. But what they've done is they've almost they they've used their system models which are one behind and they haven't changed the pace on that. So we're getting these wonderful releases with the likes of Opus and Sol for Claude and Open Chat GBT. But you know, they're stalling Fable and Astra, but we we're getting all this goodness. So, you know, we're getting improved intelligence, we're getting better performance, and we're getting markedly better prices. And it's It's where it matters. We n not everybody, most companies, the the the level of intelligence now is more than adequate on these system models for general knowledge work and computer use. And you know, they're not holding that back. And it's I just find that fascinating and wonderful. Ralph: Yeah, so it kind of you know, really begs the question just how all of this starts to go mainstream. I was on the retail innovation podcast with Craig Smith yesterday, and we were actually talking about the rise of these personal assistants. And obviously Meta has really kind of come into the news over the last few weeks with Muse. before we had the the Grokbot. prior to that, obviously Hermes agents made a a big splash. Prior to that, the whole thing kicked off with with the whole clawed bot open claw which Peter Steinberger released, and that was really the start of this concept of the dedicated personal assistant. Now Meta have joined the fray with Muse, and what was so interesting about it is that it just seems to have sparked kind of a mass adoption. And one of the reasons why seems to be because they've just been able to make it accessible to the layperson. So open claw was incredibly hard. we had all those security scares. We spoke about this when Grok released their agent, what, now probably six weeks ago. And and that personal assistant really skyrocketed and changed the fort fortunes, I think, of Grok. It drew a lot of attention. People were lauding it for its ease of use. Muse has even trumped that. And so it's kind of fascinating to see just how some of these AI solutions are actually breaking into the mainstream because we shouldn't forget, you know, we live and breathe this every day. But for the most part, when I talk to people out there, they're just kind of like, yeah, I think our company is doing something with AI. maybe they're y you know, still looking for the family holiday on Chat GPT. But there's still you know, it's still just the kind of top two percent that seem to be doing anything meaningful. But maybe now with these personal assistance, we're actually starting to get breakthrough. Em: I totally agree. And and just to give you indication of how fast this adoption has happened, it's currently the fastest growing consumer application on the market. It's even outpacing the historic launch velocity of Chat GBT. That in itself is telling. Ralph: Yeah, I mean I came across a listing that showed it was the top of the app store downloads. So that's incredible, you know, coming from nowhere. Em: Glad I'm glad you reminded me of that because I if people aren't aware, Chat GBT has actually been on the number one spot for years and it's been very infrequently that it's been knocked off its perch and it was only temporarily and when it did it went back down to two and then went back up to number one and now it's off its perch. Ralph: Yeah, and just to kind of illustrate, you know, we've got at last check, I think Anthropic were kicking off, kicking around sixty-four billion in turnover. OpenAI with its suite of tools at around 40 billion. I think now with the Sol release, they really kind of catching up. And so we've just seen a real shake shake up of the the leaderboards. and I'm you know ever so pleased to actually see That Meta have pulled something off. They have an incredible wealth of knowledge and understanding and the access to data. So if anybody is able to build consumer-facing tech, I would absolutely put my money behind Meta to do that. The fact that they haven't been able to actually, you know, you know, really bring anything to market on the software side, what is incredible as well is that they Have also launched another piece of hardware. I haven't had much time to look at it, but it's called Charm. and so this is essentially do do you remember a piece of hardware that some Chinese guy had launched? It was called Rabbit. It was a little AI on a keychain fob. anyway, basically Em: yeah, yeah. Ralph: Meta have gone down the same route and they've actually r released charm. and I think you know, we've been talking about embodied AI, ambient AI. Embodied AI is more about the kind of AI in in robots. ambient AI is more about the kind of passive device you have hanging around your your neck. And so charm is one of these devices that is supposed to put the fun into AI also from meta. So they've now got Muse as the personal assistant, they've got the meta glasses, and they've got this bit of hardware. So some really interesting developments from them. Em: Yeah, I just wanted to add, for those unaware, the charm is in the style of a Tamaguchi. It does look like fun. I haven't looked in enough about it, but that in the back end is connected to Muse. So you've got wearable ambient tech, which we covered early part of this year now, fully engaged with a consumer agent with all the simplicity and ease of use. Super cool. And we we spoil for choice. Not only do we have Muse, we had Grockbust released a week or two ago. So yeah, fill your boots, check it out. Ralph: Yeah, I really feel like that's an area that everybody is now kind of able to adopt their own personal assistant. And I was talking to a friend of mine in the last week that has been using a little device called Omni. it hangs around your neck. it's just a looks like a little kind of round pebble type thing. and it's got a a a little red dot A a light that shows that it's active and it basically just takes in all ambient conversation. And he was showing me some of the stuff that he's been able to do with that. And he's been able to make some incredible content, basically taking all of his conversations and turning them into documentaries with you know video that he's been recording. Unbelievable. So interesting for our show, Emil, because it literally means every moment of the day is an opportunity to be cutting some podcast content. And and as we know, obviously, you know, what I think everybody's responding to is the stuff that's real. This is us guys, here we are. my AI wouldn't be waving its hands around like this, because it makes you look ridiculous. But you know, the fact of the matter is we want authentic human-made content. Commentary, we want those moments where things go wrong, those things that are just funny, everything that's so polished and so clean, it just starts to become a little bit laborious and boring over time. And so I think you know, the whole ambient tech is a real opportunity because actually finding the time to do these podcasts, doing all of the prep, you know, it takes a lot of time and effort. And sometimes during the day. We're both having conversations that are worthy of a podcast. You know, we're jumping into something about a newly released bit of tech or something that we've been experimenting with. And I think those moments are sometimes even more gold than you know, what we're able to achieve on the show. So yeah, let's see. Hopefully we, you know, we'll have to get ourselves a couple of these devices and just kind of bring in, you know. it's gonna be the the hits of the week, ambient moments, that we will have to go and and share as its own special podcast release. Em: I love it. And then when it's th those that ambient tech is recording things, you can just hand it over to your agents to find the best bits of the week, our best conversations, create a real and self publish. Exciting stuff. Ralph: And it literally takes seconds. I mean, I've been watching him publish some of this stuff and he was like, Okay, yeah, I've uploaded the video, I've uploaded the and it just pairs the the the video and the images along with the the voiceover and just all you know, super possible. like super awesome. But what he's doing is kind of interesting. It's taking the transcript. So it's not actually recording audio, it's recording the transcript, which means it can fit a lot more on that small device. And then what he's doing is he's just passing it through his voice clone on eleven Em: Yeah. Ralph: eleven labs. and that's actually what's generating the video the the audio for the video. So I have to say, it sounds great. I just I again I of I I I I think after you get in small snippets it's great, but when you have to sit and listen to forty five minutes of that, maybe it's not as great. I don't know. we you know, we're probably getting to the place now where all of the mistakes and just weirdness of human communication we can kind of even program in as a bit of a profile. But yeah it's really awesome stuff. So one of the things I did want to just briefly jump into because I I did a a somewhat of a a deep dive. It's actually on the the bleeding edge podcast r resources. I had a quick look At this Soul Pi. one of the reasons I've been looking at that is because you know we've just seen that with the likes of Fable and even the likes of Soul, which is the the the flagship model from ChatGPT, that actually you can run out of tokens pretty quickly. and one of the ways that I've been combating that recently is I've actually been using a prompt generator. And this prompt generator builds the prompt for each specific model that I'm going to use. And so if I'm doing some analysis for a client with with Nexite, for example, you know, I might say, okay, let's just jump in. it's a a a defined model, a defined process. There's a a really nice README file that sets it out. So I'll use Sonnet for that. But if it's a little bit more complicated, it's I'm trying to build the model, then I might use Opus 5.5. If I'm actually putting together a dashboard, then I would probably pick Fable for that and give it a kind of a one-shot prompt. But each of them requires a very different style of prompting. And so I will typically use my prompt generator to build a prompt specifically for the model and I don't know what you found, but the differences are staggering because you know we cannot keep up with the prompt engineering changes that are required to optimize the use for each model. So I highly recommend if you haven't done this, you know, get a hold either of my particular skill, I'll figure out a way of putting it on the website, on a GitHub, you can download it, or build your own. and that has been kind of my biggest defense. But now what we're seeing with NVIDIA is. they're actually putting this soul pie model out there. And it's reported to obviously cut coding agent traffic by up to 49%. And I wanted to find out well, you know, can we actually use this? And I did a bit of deep research on this, and so I'm just gonna read a little bit of this because I actually I haven't had time to to fully assimilate it, but this is definitely something I'm gonna play with. So so SaltPi does not wrap your coding agent in a research loop. It it's the auto research loop is the method that NVIDIA used to discover the mechanisms. So agents proposed 152 candidate efficiency ideas. It tested them on 535 executable environments and threw away every idea that cost capability. The product is what survived. Four mechanisms shipped as an MIT licensed. TypeScript extension for Pi, the open source coding agent, formerly developed as BadLogic Pi-Mono. Anyway, I'd I've been looking at Pi. so it's a coding harness. so that's kind of interesting. I'm guessing in that case, it's actually not going to be able to be used with Claude. but it's just a fascinating way that they've kind of, you know, I want to use the word distilled. But you know, looking at 152 candidate efficiency ideas, testing them on five hundred and thirty-five executable environments, and then basically distilling from that what really worked. And that is what they then put into this particular model that Nvidia have open sourced. So so what's actually in the box, there's I don't know if this is gonna mean anything to us, but they've put something together called Action Fusion. it basically removes the model turn between edit a file and run the tests. So what it does is it replaces Pi's edit and write tools with versions that take an optional then run command. So the harness applies the edits, checks with a hash that nothing else is changed in the file, runs the command and returns one combined result. Anyway, probably a little bit too much detail for the show, but they've also got observation pack, evidence preserving reducer. online context compact and the agent keeps a plan through an update plan tool and basically it means that each completed step is a candidate point for compaction so those four steps essentially those four features in the end give you that 49% reduction and it looks like it only works with pie so if you are a pie user pie is a a harness a coding harness that Is one of these tools that you can really tailor specific to your workflow. The benefit of it is you control it. the difficulty that we're all seeing with the likes of Claude and Codex is they're changing every week, and sometimes things are breaking. And so the benefit of using Pi is that you can actually set it up in a custom way, you know how it works. So some interesting stuff there with with Sol. Pi, I wish it could work on on clawed code on the clawed code harness. But of course, you can push GPT and clawed models through your Pi harness. So this isn't like it's not accessible to those coding agents. You just need to be setting it up within the the Pi harness. So maybe that's a good reason to think about switching, Emil. I don't know. Are you brave enough to to move to Pi? Em: I'm very excited. I definitely wanna try it out. It's more about just finding available time with all these wonderful products that are dropping. But I'm glad you raised that point because while all these really cool, interesting, helpful products are coming out, the labs themselves are also working hard on bringing that intelligence silencly silently under the bonnet to to their own models. And by way of example, there's there's two particular things, there's adaptive compute. And adaptive reasoning and what that effectively means is when you're prompting it, it's working out how much reasoning and which model it should use. So that's gonna be welcomed. I think where I've got an open question is in some instances, as you become an expert, you want a little bit more control over that. So I'm not entirely sure how that's gonna roll out. but from what I understand, both both the major labs are actively trying to kill the manual drop-down menu. Instead of forcing a user or developer to choose between premium or mini variations. So yeah. Ralph: Well, I mean and a and even the setting, right? Because you want to run it on low, medium, high, extra high, ultra. Em: Exactly. Ralph: and actually I saw a post from a a user this week that said he ran everything on low and it worked brilliantly. If something was stuck, he didn't you know, increase the the amount of thinking it did from medium to high or extra high. He actually just flipped to a better model. and he said he was basically using about eight percent of his tokens compared to when he used Claude or even you know, the Chat GPT models on high. and he recommended basically saying, well, forget about it, you know, don't don't worry about that setting. Just use everything on low. it does the same work. A couple of times you'll find that maybe it doesn't solve the problem, just bump up the model, keep it on low, and you are using less than ten percent of the tokens. Em: That's really good advice and it's so funny you mentioned that because I was having such an excellent results session in the last two days developing with Opus five point five on medium. I started getting a little bit arrogant and confident thinking, you know, I don't need it on medium. I s in in quite a few instances I switched it to low and it one shot at it, but not a problem. Ralph: Yeah, so maybe this is the secret, you know, it just gets super, super expensive. My own kind of default has been you know, especially when I'm crunching a lot of data, which I typically do with with the next site team, you know, we are often running a lot of deep analysis on some very big retail data sets. And I I typically will run it on high or even extra high at times. but I've definitely started to tone it down a little bit. and actually, you know, w yeah, I'm kind of more in favor of doing multiple passes as opposed to one shotting it. And I think this has kind of been one of the the interesting things that's changed for me in terms of how I've been working with the tools over the last while. Well one of the other things I've also done, I've mentioned this in the past, is I've been using codex. as a code review. And so you know other people might do it a different way they use codex to code and they use claude as the code review. I've been using Claude as my my main agent. It's what we deploy also at at Nextite. but of course you know actually having Codex do the code review but one of the things I found is when I run out of like Codex tokens, then I switched to Kemi. And and my the most fascinating learning there, Emil, has been I've been it it kind of says, look, the this Kimi has got blind spots. And so you know obviously there's a prompt that tries to to deal with Kimi's blind spots. but the most amazing thing is I've now actually got three models looking at the same problem and each of them are actually finding different bugs, finding different ways to execute better. And I just I feel like it's it's improved the process drastically and it's given me a lot more confidence that what I'm putting out there has been peer reviewed from multiple angles. And I don't know. I mean it it's obviously an awful lot of processing that's happening. but at the same time just starting to actually leverage the different models. for their strengths. I think what it's starting to show is that you know you can't live in a one model world you really need to kind of open up the your your capability and use different models for for different parts of the solution code review design review testing all of that works very very well with the likes of codex and chemi in partnership with with claude code Em: That's really good at great advice, Rolf. And I just want to restress what you've already said, that that whole point on higher intelligence is certainly try work on the default level and think of high intelligence rather as cracking a nut when you're in a jam and you get completely stuck. It's kind of like your get out of jail not so free card and then, you know, once you get around that corner, switch back because I I welcome anybody who's sincerely interested in this. If you look up the the graphs on the potential on all the models and you look at the price points and the increase of intelligence, the i increase in intelligence as you step up is quite marginal, but the price point goes hyperbolic. So you're paying a lot more for getting very little incremental increases in intelligence. So maybe think about it that way when you code or develop a or use your And knowledge work. It's really helpful. Thanks for that, Rolf. Ralph: Yeah, and of course, you know, if you don't actually use multiple models, you can even just do this within Codex if you're on the OpenAI stack or if you're on Claude Code, you can literally s you know just spawn another terminal and tell that terminal you are a reviewer, you can pick a cheaper model, you can pick an expensive model, and one of the things you can do if you hover over the terminal name. you do have to set this up, but you've got internal messaging between terminals. I spoke about this a couple of months back because it was a real revelation for me. getting the terminals to communicate with each other. and I stumbled on this solution about three weeks ago, where I had two terminals working on a similar piece of work, and there was some overlap starting to happen. And so I actually said to the one, you need to tell the other terminal what you're busy working on. And they started to correct each other. So they were actually then performing the review. but there wasn't like I had a review terminal. They both became the reviewer for each other. So this was like having you know, back in the old days, you would have a code buddy, right? So somebody that reviewed your code, the terminals became They're code buddies and both of them had context about the same thing, looking at slightly different problems. but I really recommend kind of setting up a bunch of terminals, having them work independently on a specific aspect, giving them some shared context, and then letting them actually review each other's work from somewhat different points of view. Incredibly powerful, even just within the same model. Em: That sounds amazing. I've I haven't come across that. I'm quite keen to do it, but I wouldn't know how to get started. How how exactly do you create the bridge for them to speak to each other? Ralph: well, I it's called there's a special term for it, but it's it's called messaging. I've got a a setup instruction for it which I'll make available on the show notes. and you it's a little bit of work, you have to basically install it and get them talking to each other. but when they do, you can just invoke it at any point in time. And I regularly get terminals just to share what they're working on. To make sure that they're not stepping on other terminals' toes. And also when you're reaching context, this is a really great thing. When you when you're just kind of like, you know, you're at 90%, you know you're gonna have to shut that terminal down. What I'll typically do now is I'll say, right, let's start the window process. Part of the window process is you will generate a prompt for a brand new, fresh sibling terminal. To pick up the activity. So generate the prompt, give it access to everything that you think is outstanding and and so you know, give it access to all of the context it needs. and I then paste that into the new terminal. I then tell the new terminal to go in and do a a check with the the terminal that is has just handed over. And ask it any clarifying questions, anything it doesn't understand from the briefing, any additional context it wants. And this has been absolutely incredible because what I'm now getting is I'm preserving all of that context from you know what sometimes is a terminal session that might run for two weeks. and as I'm concluding a piece of work, there's still a little bit more to do. I'm handing over to a fresh terminal and it gets all of the briefing. And the opportunity to actually clarify the meaning. And so it's the cheapest way of building context. And it's what been one of the best handoff procedures that I've come across. And so I can seamlessly move and transition from an outgoing terminal that's about to reach, you know, 100% context and just move into a fresh terminal that is then able to tackle the problem with 100% context or, you know, however much you've got available. and that's just been brilliant. Em: It's amazing. More fun, time consuming homework to enjoy. Ralph: Well, just make sure it's productive. I think this is kind of, you know, sometimes when you're dealing with a difficult problem that you're trying to solve, it's it's just good to kind of be able to move across terminals and not feel like, I've kind of hit the cliff. you know, I have to shut this down, wrap it up. it's actually just a a really good point to just gather your thoughts, consolidate, get to the nub of what is it that I'm trying to solve, brief the next terminal on that. and it just stops the drift. I think this is the the the reason I kind of came up with this approach. was actually down to opus because it was just so laborious and and it was so prone to drift that I needed to kind of focus down and get back to well what are we actually trying to solve? Like I'd spent you know whole sessions where it would just literally navel gaze deep into its dark, dark soul to come back with absolute gibberish. And and not actually achieve anything. I remember sitting a couple of Saturdays ago, literally running prompts all day trying to achieve something on the the music pipeline, and this thing just w went into a navel gazing you know like doom loop that didn't come out for eight hours, and so I started to kind of go, okay, no, you go let's finish this session up. Let's refocus on what we're trying to do. Let's hand it over to a new terminal. so I guess that's one good thing that came out of Opus V. Em: Well, I think that's a wrap. We're on the hour and I'd just like to thank our listeners. please do like and subscribe. It really helps us out. to bring the show to you. We enjoy it and we wanna continue. Ralph: Yeah, it's been an amazing journey so far. It's incredible to see so many new subscribers and just the the interest in this. I mean we're really only just getting going in terms of the mainstream. and with so much going on, I think you've got to find a way to really keep up with the news. And really that's what this channel is all about. So bringing you the bleeding edge. it was difficult in the past. I had to find all these innovators and people that were doing amazing things that hadn't yet kind of come to the mainstream. But now, you know, with AI every week from at least, you know, a hundred different companies, we've got evolutions that are absolutely at the bleeding edge, brand new technology and capability. And what's incredible about this moment is it's not just for the early adopters and those people at the fringe in the know. This is technology that is available to everybody, which is a staggering. So I really feel like the bleeding edge in its current form of kind of AI news reporting is one of the biggest gaps that we can fill in the market in terms of information. So it's a really a real delight to be able to share that with everybody. Em: Absolutely. Jevon's paradox playing through and coming a r reality. Ralph: And the probability of a show happening next week, Emil, has Jeev calculated what that's likely to be? Em: I've just asked him, he said a hundred percent. Ralph: There we go. And with that, we'll see you on the edge. Em: Thanks everyone.